CVE-2026-33235 is a high-severity vulnerability in Significant-Gravitas AutoGPT prior to version 0.6.52. It involves uncontrolled resource consumption (CWE-400) in the Fill Text Template block. Although the backend SandboxedEnvironment prevents unauthorized attribute access, it fails to constrain the computational complexity or execution time of user-supplied expressions. Attackers can craft computationally expensive Python or Jinja2 expressions that consume excessive CPU and memory resources, resulting in denial of service through system hangs or crashes. This vulnerability affects multi-tenant and self-hosted environments, causing service outages and noisy neighbor effects that require manual intervention. The vulnerability has been addressed and fixed in AutoGPT version 0.6.52.

Exploitation of this vulnerability leads to denial of service by exhausting CPU and memory resources on the server running AutoGPT. This causes the system to hang or crash, resulting in complete service outages. In multi-tenant or self-hosted environments, this can also cause noisy neighbor effects, impacting other tenants or services on the same infrastructure. There is no impact on confidentiality or integrity reported.

Upgrade AutoGPT to version 0.6.52 or later, where this vulnerability has been fixed. The fix involves limiting the computational complexity and execution time of expressions in the Fill Text Template block. Until upgraded, administrators should consider restricting access to the vulnerable functionality or monitoring for signs of resource exhaustion. Patch status is confirmed fixed in version 0.6.52.